Attention-Enhanced Random Forest for Power Output Estimation of a 660 MW Supercritical Thermal Power Unit

The increasing integration of renewable energy into power grids requires thermal power units to operate with enhanced flexibility, and accurate power output estimation is a fundamental task for achieving flexible operation. A power output estimation method integrating an AM (Attention Mechanism) with a RF (Random Forest) is proposed for thermal power units. The attention mechanism identifies key features from 20 candidate operational parameters through a two-stage feature selection process. In the first stage, Pearson correlation coefficient screening retains 16 parameters. In the second stage, attention weights are computed to rank the retained parameters by importance, and the top-ranked features are selected as the final input to the random forest model. Based on DCS (Distributed Control System) historical operation data from a 660 MW supercritical Unit 2, the proposed method is systematically compared with six mainstream predictive models under a unified evaluation framework. The results show that the base random forest model of the proposed method achieves the lowest mean absolute error of 32.90 MW, a coefficient of determination of 0.83, and a computation time of 25 min among all seven models. After introducing the two-stage feature selection of the attention mechanism, the complete AM-RF (Attention Mechanism-Random Forest) model reduces the error by 56% compared with the long short-term memory network and by 75% compared with the autoregressive moving average model. The coefficient of determination reaches 0.98, and the computation time is shortened by 23% and 13%, respectively. In the linear fitting validation across three load intervals, the fitting slopes are all close to 1, and the coefficients of determination remain consistently high. The synergy mechanism between the attention mechanism and the random forest effectively improves the estimation accuracy, robustness, and computational efficiency for power output estimation of thermal power units.